Hi Dean,
Does this mean I am going to add a
new object into M5(write C++ code for perfect_branch_predictor, then
instantiate it using python)?
I'm thinking you would have to do write a "perfect_branch_predictor"
object to plug in.
The only complication is you have to figure out how to resolve the
branches of speculative instructions if you are using that for an
out-of-order machine.
If your workload is small enough, an idea may be figuring out how to
pre-execute your workload and save the the branch history so as to use
as input to your perfect predictor.
There is no way to do a perfect branch predictor without future
knowledge of instructions. This is not that easy to do with our current
CPU model that executes instructions at execute (a coupled model, as
opposed to a functional first model). A possible option is like what
Korey mentioned, to pre-execute the workload using a simple CPU model
and save the branch history for future use. Another option is to
instantiate another system and use its instruction stream to determine
branch directions, but that's a bit more involved.
Kevin
_______________________________________________
m5-users mailing list
m5-users@m5sim.org
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
